Yet another backpropagation tutorial

I am teaching deep learing this week in Harvard’s CS 182 (Artificial Intelligence) course. As I’m preparing the back-propagation lecture, Preetum Nakkiran told me about Andrej Karpathy’s awesome micrograd package which implements automatic differentiation for scalar variables in very few lines of code.

I couldn’t resist using this to show how simple back-propagation and stochastic gradient descents are. To make sure we leave nothing “under the hood” we will not import anything from the package but rather only copy paste the few things we need. I hope that the text below is generally accessible to anyone familiar with partial derivatives. See this colab notebook for all the code in this tutorial. In particular, aside from libraries for plotting and copy pasting a few dozen lines from Karpathy this code uses absolutely no libraries (no numpy, no pytorch, etc..) and can train (slowly..) neural networks using stochastic gradient descent. (This notebook builds the code more incrementally.)

Automatic differentiation is a mechanism that allows you to write a Python functions such as

def f(x,y): return (x+y)+x**3

and enables one to automatically obtain the partial derivatives \tfrac{\partial f}{\partial \text{x}} and \tfrac{\partial f}{\partial \text{y}}. Numerically we could do this by choosing some small value \delta and computing both \tfrac{f(x+\delta,y)-f(x,y)}{\delta} and \tfrac{f(x,y+\delta)-f(x,y)}{\delta}.
However, if we generalize this approach to n variables, we get an algorithm that requires roughly n evaluations of f. Back-propagation enables computing all of the partial derivatives at only constant overhead over the cost of a single evaluation of f.

Back propagation and the chain rule

Back-propagation is a direct implication of the multi-variate chain rule. Let’s illustrate this for the case of two variables. Suppose that v,w: \mathbb{R} \rightarrow \mathbb{R} and z:\mathbb{R}^2 \rightarrow \mathbb{R} are differentiable functions, and define

f(u) = z(v(u),w(u)).

That is, we have the following situation:

where f(u) is the value z=z(v(u),w(u))

Then the chain rule states that

\tfrac{\partial f}{\partial u} = ( \tfrac{\partial v}{\partial u} \cdot \tfrac{\partial z}{\partial v} + \tfrac{\partial w}{\partial u} \cdot \tfrac{\partial z}{\partial w} )

You can take this on faith, but it also has a simple proof. To see the intuition, note that for small \delta, v(u+\delta) \approx v(u) + \delta \tfrac{\partial v}{\partial u}(u) and w(u+\delta) \approx w(u) + \delta \tfrac{\partial w}{\partial u}(u). For small \delta_1,\delta_2, z(v+\delta_1,w+\delta_2) \approx z(v,w) + \delta_1 \tfrac{\partial z}{\partial v}(v,w) + \delta_2 \tfrac{\partial z}{\partial w}(v,w). Hence, if we ignore terms with powers of delta two or higher,

f(u +\delta)= z(w(u+\delta),v(u+\delta)) \approx f(u) + \delta \tfrac{\partial v}{\partial u} \cdot \tfrac{\partial z}{\partial v} + \delta \tfrac{\partial w}{\partial u} \cdot \tfrac{\partial z}{\partial w}

Meaning that \frac{f(u +\delta) - f(u)}{\delta} \approx \tfrac{\partial v}{\partial u} \cdot \tfrac{\partial z}{\partial v} + \tfrac{\partial w}{\partial u} \cdot \tfrac{\partial z}{\partial w} which is what we needed to show.

The chain rule generalizes naturally to the case that z is a function of more variables than u. Generally, if the value f(u) is obtained by first computing some intermediate values v_1,\ldots,v_k from u and then computing z in some arbitrary way from v_1,\ldots,v_k, then \tfrac{\partial z}{\partial u} \sum_{i=1}^k \tfrac{\partial v_i}{\partial u} \cdot \tfrac{\partial z}{\partial v_i}.

As a corollary, if you already managed to compute the values \tfrac{\partial z}{\partial v_1},\ldots, \tfrac{\partial z}{\partial v_k}, and you kept track of the way that v_1,\ldots,v_k were obtained from u, then you can compute \tfrac{\partial z}{\partial u}.

This suggests a simple recursive algorithm by which you compute the derivative of the final value z with respect to an intermediate value w in the computation using recursive calls to compute the values \tfrac{\partial z}{\partial w'} for all the values w' that were directly computed from w. Back propagation is this algorithm.

Computing \tfrac{\partial z}{\partial u},\tfrac{\partial z}{\partial v}, \tfrac{\partial z}{\partial w} for the assignment u=5 using back propagation

Implementing automatic differentiation using back propagation in Python

We now describe how to do this in Python, following Karpathy’s code. The basic class we use is Value. Every member u of Value is a container that holds:

  1. The actual scalar (i.e., floating point) value that u holds. We call this data.
  2. The gradient of u with respect to some future unknown value f that will use it. We call this grad and it is initialized to zero.
  3. Pointers to all the values that were used in the computation of u. We call this _prev
  4. The method that adds (using the current value of u and other values) the contribution of u to the gradient of all its previous values v to their gradients. We call this function _backward. Specifically, at the time we call _backward we assume that u.grad already contains \tfrac{\partial z}{\partial u} where z is the final value we are interested in. For every value v that was used to compute u, we add to v.grad the quantity \tfrac{\partial z}{\partial u} \cdot \tfrac{\partial u}{\partial v}. For the latter quantity we need to keep track of how u was computed from b.
  5. If we call the method backwards (without an underscore) on a variable u then this will compute the derivative of u with respect to v for all values v that were used in the computation of u. We do this by applying _backward to u and then recursively (just like in DFS) going over the “children” (values used to compute u), calling _backward on each one and keeping track the ones we visited just like the Depth First Search (DFS) algorithm.

Let’s now describe this in code. We start off with a simple version that only supports addition and multiplication. The constructor for the class is the following:

class Value:
    """ stores a single scalar value and its gradient """

    def __init__(self, data, _children=()): = data
        self.grad = 0
        self._backward = lambda: None
        self._prev = set(_children)

which fairly directly matches the description above. This constructor creates a value not using prior ones, which is why the _backward function is empty.
However, we can also create values by adding or multiplying prior ones, by adding the following methods:

  def __add__(self, other):
        other = other if isinstance(other, Value) else Value(other)
        out = Value( +, (self, other))

        def _backward():
            self.grad += out.grad
            other.grad += out.grad
        out._backward = _backward

        return out

    def __mul__(self, other):
        other = other if isinstance(other, Value) else Value(other)
        out = Value( *, (self, other))

        def _backward():
            self.grad += * out.grad
            other.grad += * out.grad
        out._backward = _backward

        return out

That is, if we create w by adding the values u and v, then the _backward function of w works by adding w.grad = \tfrac{\partial z}{\partial w} to both u.grad and v.grad.
If we w is obtain by multiplying u and v then we add w.grad \cdot = \tfrac{\partial z}{\partial w} v to u.grad and similarly add w.grad \cdot = \tfrac{\partial z}{\partial w} u to v.grad.

The backward function is obtained by setting the gradient of the current value to 1 and then running _backwards on all other values in reverse topological order:

def backward(self, visited= None): # slightly shorter code to fit in the blog 
    if visited is None:
        visited= set([self])
        self.grad = 1
    for child in self._prev:
        if not child in visited:

For example, if we run the following code

a = Value(5)
def f(x): return (x+2)**2 + x**3

then the values printed will be 0 and 89 since the derivative of (x+2)^2 + x^3 = x^3 + x^2 + 4x + 4 equals 3x^2 + 2x +42.

In the notebook you can see that we implement also the power function, and have some “convenience methods” (division etc..).

Linear regression using back propagation and stochastic gradient descent

In stochastic gradient descent we are given some data (x_1,y_1),\ldots,(x_n,y_n) and want to find an hypothesis h that minimizes the empirical loss L(h) = \tfrac{1}{n}\sum_{i=1}^n L(h(x_i),y_i) where L is a loss function mapping two labels y, y' to a real number. If we let L_i(h) be the i-th term of this sum, then, identifying h with the parameters (i.e., real numbers) that specify it, stochastic gradient descent is the following algorithm:

  1. Set h to be a random vector. Set \eta to be some small number (e.g., \eta = 0.1)
  2. For t \in {1,\ldots, T} (where T is the number of epochs):
  • For i \in {1,\ldots, n}: (in random order)
    • Let h \leftarrow h - \eta \nabla_h L_i(h)

If h is specified by the parameters h_1,\ldots,h_k \nabla_h L_i(h) is the vector ( \tfrac{\partial L_i}{\partial h_1}(h), \tfrac{\partial L_i}{\partial h_2}(h),\ldots, \tfrac{\partial L_i}{\partial h_k}(h)). This is exactly the vector we can obtain using back propagation.

For example, if we want a linear model, we can use (a,b) as our parameters and the function will be x \mapsto a\cdot + b. We can generate random points X,Y as follows:

Now we can define a linear model as follows:

class Linear:
  def __init__(self):
    self.a,self.b = Value(random.random()),Value(random.random())
  def __call__(self,x): return self.a*x+self.b

  def zero_grad(self):
    self.a.grad, self.b.grad = 0,0

And train it directly by using SGD:

η = 0.03, epochs = 20
for t in range(epochs):
  for x,y in zip(X,Y):
    loss = (model(x)-y)**2
    model.a , model.b = (model.a - η*model.a.grad  , model.b - η*model.b.grad)

Which as you can see works very well:

From linear classifiers to Neural Networks.

The above was somewhat of an “overkill” for linear models, but the beautify of automatic differentiation is that we can easily use more complex computation.

We can follow Karpathy’s demo and us the same approach to train a neural network.

We will use a neural network that takes two inputs and has two hidden layers of width 16. A neuron that takes input x_1,\ldots,x_k will apply the ReLU function (max{0,x}) to \sum w_i x_i where w_1,\ldots,w_k are its weight parameters. (It’s easy to add support for relu for our Value class. Also we won’t have a bias term in this example.)

The code for this Neural Network is as follows: (when Value() is called without a parameter the value is random number in [-1,1])

def Neuron(weights,inputs, relu =True):
  # Evaluate neuron with given weights on given inputs
  v =  sum(weights[i]*x for i,x in enumerate(inputs))
  return v.relu() if relu else v

class Net:
  # Depth 3 fully connected neural net with one two inputs and output
  def __init__(self,  N=16):
    self.layer_1 = [[Value(),Value()] for i in range(N)]
    self.layer_2 = [ [Value() for j in range(N)] for i in range(N)]
    self.output =  [ Value() for i in range(N)]
    self.parameters = [v for L in [self.layer_1,self.layer_2,[self.output]] for w in L for v in w]

  def __call__(self,x):
    layer_1_vals = [Neuron(w,x) for w in self.layer_1]
    layer_2_vals = [Neuron(w,layer_1_vals) for w in self.layer_2]
    return Neuron(self.output,layer_2_vals,relu=False) 
    # the last output does not have the ReLU on top

  def zero_grad(self):
    for p in self.parameters:

We can train it in the same way as above.
We will follow Karpathy and train it to classify the following points:

The training code is very similar, with the following differences:

  • Instead of the square loss, we use the function L(y,y')= \max{ 1- y\cdot y', 0 } which is 0 if y \cdot y' \geq 1. This makes sense since our data labels will be \pm 1 and we say we classify correctly if we get the same sign. We get zero loss if we classify correctly all samples with a margin of at least 1.
  • Instead of stochastic gradient descent we will do standard gradient descent, using all the datapoints before taking a gradient step. The optimal for neural networks is actually often something in the middle – batch gradient descent where we take a batch of samples and perform the gradient over them.

The resulting code is the following:

for t in range(epochs):
  loss = sum([(1+ -y*model(x)).relu() for (x,y) in zip(X,Y)])/len(X)
  for p in model.parameters: -= η*p.grad

If we use this, we get a decent approximation for this training set (see image below). As Karpathy shows, by adjusting the learning rate and using regularization, one can in fact get 100% accuracy.

Update 11/30: Thanks to Gollamudi Tarakaram for pointing out a typo in a previous version.

2 thoughts on “Yet another backpropagation tutorial

  1. The magic of backpropragation is that it computes all partial derivatives in time proportional to the network size. This is much more efficient than computing derivatives in “forward mode”.
    So backprop not only works in practice,it also works in theory! For a reason that I do not understand, researchers on neural networks overemphasized the first point (practical efficiency) to the detriment of the second…

  2. Yes – as mentioned the back propagation algorithm requires only one network evaluation as opposed to n (where n is the number of weights which is basically the size of the network). When I teach introduction to theoretical computer science I often use this as an example of how the difference between an asympotatically quadratic and linear algorithm makes huge difference in practice.

Comments are closed.